Poster Presentation

Poster D in Poster Session D: Wednesday, August 5, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Structure, disorder, and dynamics in task-trained recurrent neural circuits

David G Clark¹, Blake Bordelon¹, Jacob A Zavatone-Veth¹, Cengiz Pehlevan¹; ¹Harvard University

Presenter: David G Clark

Neurons in many brain areas produce heterogeneous, seemingly disordered responses during behaviors, yet the circuits they comprise must contain enough structure to support the representations and computations underlying these behaviors. How much learned structure is present in recurrent connectivity relative to disorder, and how the interaction between them shapes dynamics and single-neuron responses, remain open questions. Recurrent neural networks (RNNs) trained to perform tasks have become a leading class of models for such circuits, but conventional training yields a single point in a vast space of task-compatible solutions, with no systematic way to explore this space and no theory of how internal representations vary within it. Here, we introduce a control parameter γ that governs the degree to which learning reshapes recurrent connectivity, interpolating between a reservoir regime and one in which recurrent weights are restructured to produce task-relevant internal representations. Varying γ generates a family of task-compatible solutions whose internal dynamics differ in a controlled way. We derive a dynamical mean-field theory (DMFT) showing how the balance of randomness and structure influences population-level dynamics and single-neuron responses. In the reservoir limit, the single-neuron response distribution is Gaussian; recurrent restructuring drives it toward task-dependent, non-Gaussian forms. In nonlinear networks, restructuring drives a phase transition from chaotic, high-dimensional activity to ordered, low-dimensional dynamics that generalize temporally beyond the training period. We apply this framework to a reaching task in which an RNN must reproduce macaque muscle activity, and find that optimally matching simultaneous motor-cortex recordings requires an intermediate degree of restructuring in which learned structure coexists with random heterogeneity.

Topic Area: Methods, Tools, Theory & Neural Coding